Shape spaces and metrics in an application perspective
Mads Nielsen
eScience Centre, DIKU, University of Copenhagen
Overview
The application perspective:
Data and task
Which structure is needed on the models to subserve the
task?
Disclaimer
Slides have been stolen from Mumford,
Dam, Chennai, and many more
Osteoarthritis (OA)
OA is a degenerative joint disease in knees, hips, … Effect:
Pain, Reduced range of motion
Rule of thumb:
Age in years gives % chance of OA
Treatment:
Symptom control
Current golden standard:
-Kellgren & Lawrence Index -Joint space width
www.chclibrary.org www.chclibrary.org
Quantification Framework
Folkesson, Dam et al. 2007 Trans Medical Imaging Dam et al. 2008
Medical Image Analysis Dam, Folkesson et al. 2007
Osteoarthritis & Cartilage Qazi, Dam, Karsdal et al. 2007
Osteoarthritis & Cartilage
Volume Thickness Curvature
Smoothness,
Homogeneity
Risk of vertebral fractures
Current standard of fracture grading:
Bone Mineral Density based on dual x-ray Our approach:
Statistical shape analysis
Visualization
Green: Likely to stay intact
Blue: Mean spine shape
Red: Likely to fracture
Tasks
Classification S —> [L 1 ,L 2 ,…L n ] Shape regression S(t): R n -> S
Marker regression t(S): S -> R
Prior for segmentation p(S): dist. on S In all cases, a metric on the space S of
shapes S is essential
Finding usefull metrics is non-trivial
Shapes
Shape = Geometry \ Position
Shape is a qouotient manifold (mayby embedded in Geometry space)
Metric on the geometry space, may be inherited (projected) to the shape space
Kendall : Points in R
2n\ Similarity
Start with a fixed curve parametrized by ( ) Define a local chart near :
( ) ( ) ( ). ( ), ( ) unit normal to ,
image of
a
a a
C
s s
s s a s n s
n s C
C
a r r
f f
y f
y
Î
= +
=
=
S
The set Σ of all smooth plane curves
forms a manifold!
Think of Σ geometrically
• A curve on Σ is a warping of one shape to another.
• On Σ , the set of ellipses forms a surface:
• The geometric heat equation:
is a vector field on Σ ∂ C t
t= κ
Ctn
Ct,
∂ r
Advantages of L 2 metrics
Have simple notion of a gradient to form flows
Have a beautiful theory of locally unique geodesics, thus a warping of one shape to another.
Can define the Riemannian curvature tensor. If non- positive, have a good theory of means.
Can expect a theory of diffusion, of Brownian
motion, hence Gaussian-type measures and their
mixtures.
A geodesic in the simple L 2 metric
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
θ axis
t axis
The line on the bottom is moved to the line on the top by growing
“teeth” upwards and then shrinking them again.
But distances collapse in this metric:
Fixes of this bug:
Michor + Mumford: L 2 + curvature Yezzi: Sobolev metric
Charpiat: Bounded curvature
Sommer: Finite bandwith by resampling Trouve, Younes: Diffeomorphic
Others use Hausdorff metric
Michor+Mumford
For small shapes, curvature is negative and the path nearly goes back to the circle (= the ‘origin’).
Angle sum = 102 degrees.
For large shapes, curvature is positive, 2 protrusions grow while 2 shrink. Angle sum = 207
degrees.
Hence construction depends on scale
2 2 2